Philippe E. Tissot

This study uses a 16‐yr Tropical Rainfall Measuring Mission (TRMM) Convective Features (CFs) and ERA‐Interim reanalysis data to investigate the relative importance of four large‐scale environmental variables to thunderstorms with random forest models. These four variables include Convective Available Potential Energy (CAPE), Convective Inhibition (CIN), low‐level wind shear, and warm cloud depth (WCD). First, these selected four environmental variables show a distinguished difference between CFs with and without lightning flashes. Specifically, CFs with at least one flash have higher CAPE, CIN, and lower WCD than those without lightning. Then, using these four variables, the geographical distribution of thunderstorms, especially the land‐ocean contrast in the occurrence of thunderstorms, is closely reproduced with a global random forest model. Such results suggest that a random forest model with key large‐scale environmental variables can be a useful tool to estimate the occurrence of global lightning thunderstorms. The study also investigates the relative importance of the selected variables to the occurrence of thunderstorms regionally. Relatively higher skill scores in the regional random forest model than the global one indicate the variation of roles of large scale environment variables over different regions. Though the data‐driven models can be utilized to estimate the occurrence of global thunderstorms, how to link the regional relative importance of these variable to the physical processes of thunderstorms needs further investigation.